Skip to content

Table Review and Later Understanding

Status: Review rail implemented; later structure and Understand work remain ordered next steps

Authority: Current contract (2026-08-12)

Last reviewed: 2026-08-14

Question

What happens after a table proposal is verified?

Short answer

A human reviews the region. Later stages recover finite separators and cells. Cached OCR is assigned to the selected table. Understand assigns table type, headers, values, and title-block fields.

Human decisions

Review actions include accept, adjust, reject, and add missing. Accept can save a corrected polygon without erasing proposal history. Reject records why the box is not table truth. Add missing records a human-created region that the providers missed. All of these decisions are append-only evidence.

Corrected and missing regions

Adjusted boundaries become accepted truth for later stages. Missing tables become accepted truth with a human origin label. Accepted table and title-block regions can later exclude Fast Dimension Lines ink. Review history stays available beside the corrected result.

Review-rail baselines

The review rail can report corpus baselines from completed sheets. The baselines include accepted tables, adjusted boundaries, rejected proposals, and manually added missing tables. The baselines also include rejected and missed counts per sheet. These counts are practical review proxies. These counts are not formal precision or recall.

Structure recovery and OCR assignment

After an accepted or corrected region exists, a structure pass finds finite separators and cells. Cached OCR then assigns readings to the selected table. IBM TableFormer can recover structure from one selected crop only. TableFormer does not replace whole-sheet region detection. Structure recovery and cell OCR remain later ordered work where not yet complete on the primary reviewed-region path.

Understand

Understand assigns:

  • table type
  • headers
  • values
  • title-block fields

Understand does not turn Detection provider scores into validated truth by itself.

Original heuristics retained as comparison

Original morphology, repeated rules, and cell clustering remain visible through the Geometry detector path. That history explains false positives and why proposal plus shared verification replaced a single heuristic claim.

Open Questions

  • When does cell-text accuracy measurement begin on the first fully reviewed corpus?

Sources